The Integration Of Prior Knowledge Into Medical Image Segmentation Algorithms

With the diversity of modern medical imaging technique, how to acquire interested organfrom enormous image data is a concerned topic. Computer assisted diagnosis becomes a trenddue to limitation from human operation and the rapid development of computer imageprocessing skill. Among that image segmentation is the fundamental and critical step for thefollowing operations. The challenges include time-varying shape of organ, also low resolution,noise and motion existed in the image. This paper will focus on processing two kinds of medicalapplications with incorporating shape knowledge in common. Major works are as follows:1. High-rate of disease drives cardiac research as an important field. In clinical, left ventricleboundaries is used to calculate parameters as volume and mass. Thus, aim to obtain the inner andouter boundary of left ventricle from cardiac CT image, this paper will base on active shapemodel which acquires mean shape and gray-level information from sample data and then applieson new image. In specific, traditional method rely on derivative profile to search new position;here will combine a descriptor called DAISY to construct a feature model which will takereplace of previous gray-level model, to better realize the fitting process. Meanwhile, some othersides of tradition method are adjusted, such as landmarks setting and model initialization. Inconclusion, the method afterward improves the speed, accuracy and feasibility to some degree.2. The stenosis of lumen cannot prove the arteriosclerosis degree accurately and nowadaysmulti-sequence MRI carotid image is used widely to assess the plaque vulnerability. The firstprocedure is obtaining the edge information of internal and external wall. The inner part iscomparably easier to get while the outer one is still a hard task. This paper will use thesegmentation result from internal wall as the shape prior to build a ring-like prior model, join inRSF model to evolve. The outer boundary is found when the formulation energy reaches aminimum. With a comparison with several other level-set based active contour models, thismethod seems much more appropriate.Though the job above, it is showed that incorporating some prior knowledge will decreasethe segmentation difficulty and achieve better results when the interested shape is complex orlocated in a disturbing background.